A Human-machine Collaborative Framework for Evaluating Malevolence in Dialogues

Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 5612–5623 August 1–6, 2021. ©2021 Association for Computational Linguistics 5612 A Human-machine Collaborative Framework for Evaluating Malevolence in Dialogues Yangjun Zhang University of Amsterdam y.zhang6@uva.nl Pengjie Ren∗ Shandong University renpengjie@sdu.edu.cn Maarten de Rijke University of Amsterdam & Ahold Delhaize m.derijke@uva.nl Abstract Conversational dialogue systems (CDSs) are hard to evaluate due to the complexity of nat- ural language. Automatic evaluation of di- alogues often shows insufficient correlation with human judgements. Human evaluation is reliable but labor-intensive. We introduce a human-machine collaborative framework, HMCEval, that can guarantee reliability of the evaluation outcomes with reduced human ef- fort. HMCEval casts dialogue evaluation as a sample assignment problem, where we need to decide to assign a sample to a human or a machine for evaluation. HMCEval includes a model confidence estimation module to esti- mate the confidence of the predicted sample as- signment, and a human effort estimation mod- ule to estimate the human effort should the sample be assigned to human evaluation, as well as a sample assignment execution mod- ule that finds the optimum assignment solu- tion based on the estimated confidence and ef- fort. We assess the performance of HMCEval on the task of evaluating malevolence in di- alogues. The experimental results show that HMCEval achieves around 99% evaluation ac- curacy with half of the human effort spared, showing that HMCEval provides reliable eval- uation outcomes while reducing human effort by a large amount. 1 Introduction Conversational dialogue systems (CDSs) are often trained to generate responses given unstructured, open-domain dialogues. Evaluation of CDS re- sponses has drawn broad attention due to its cru- cial rule for CDS development (Deriu et al., 2020). Broadly speaking, there are two approaches to per- form dialogue evaluation: automatic evaluation and human judgements (Finch and Choi, 2020). Automatic evaluation metrics such as appropri- ateness (Lowe et al., 2017), engagement (Zhang ∗ Corresponding author. et al., 2020), are efficient but have low agree- ment with human judgements due to the diver- sity of responses (Liu et al., 2016), especially for word-overlap based metrics, such as BLEU (Pa- pineni et al., 2002) and ROUGE (Lin and Hovy, 2002). More recently, training based methods, e.g., ADEM (Lowe et al., 2017), RUBER (Tao et al., 2018) and contextualized methods, e.g. BERT- based RUBER (Ghazarian et al., 2019), have been shown to have better agreement with human judge- ments. However, these methods are still not reli- able enough: the Pearson correlation with human judgments is 0.44 for appropriateness (Lowe et al., 2017) and 0.55 for relevance (Ghazarian et al., 2019). To guarantee reliability of evaluation out- comes, our current best practice is to use human judgements. In terms of most evaluation aspects, e.g., appropriateness (Young et al., 2018), coher- ence (Ram et al., 2018) and empathy (Rashkin et al., 2019), human judgements simply show the highest reliability. Obviously, human judgments are more labor-intensive than automatic evaluation (Deriu et …

Content truncated to 500 words